On Tue, Jan 17, 2023 at 01:13:20AM +1100, Mark Jamsek wrote:
> The below diff is borne out of previous discussions with stsp about
> eliminating the cost of computing the diff twice when displaying the
> diffstat in 'got log -d' and tog diff views.
> 
> It takes the same approach as the 'got diff -d' case where we use a temp
> file for the diff so that we can compute the diffstat with the same
> result from building the diff, and still draw the diffstat first once
> the diff is complete.
> 
> This has the additional benefit in tog (and got log -d vs got log -P -p)
> of not running the get_changed_paths() -> got_diff_tree_collect_changed_paths()
> code path to get the changeset before computing the diff (which
> traversed much the same code twice). It also results in net less code :)
> 
> From a performance perspective: at the extreme end, if you load tog with
> b5500b9ca0102f1c in OpenBSD src.git with and without this patch you will
> notice the difference. For the average diff it's not nearly as
> noticeable but is an improvement nonetheless.

That's great, thanks!

I don't have time to review the diff in detail right now, but please
proceed.

> I've also taken the liberty of making 'got log -d' behave like 'got diff
> -d' such that the actual diff is displayed too. stsp made the point when
> first reviewing got diff -d code that the diffstat per se isn't all that
> meaningful--it's most useful in the context of the diff. Regress is
> passing, but assuming this is ok, I'll follow up with a change to the
> test_log_diffstat regress to include testing for diff output. Otherwise
> I can revert this part of the diff and stick to just displaying the
> diffstat with 'got log -d', in which case the current regress provides
> sufficient coverage.

I would argue that a log message already provides good enough context
for a diffstat. A commit log message which implies a large change but
comes with a tiny-looking diffstat would stand out, and vice versa.
The 'got diff -d' case was different since it lacked context entirely.

And users are already able to quickly combine both options as "diff -dp"
if desired.

I see at least one possible use case for -d without -p style output:
Consider post-commit email notifications, where a log message with a
diffstat attached would provide a suitable amount of information, and
won't blow up peoples' inboxes when someone imports a new version of
a large vendor-imported subtree such as gcc or LLVM.
Granted, this use case is a bit contrived right now (gotd has no hook scripts,
and people running git-daemon would likely be using 'git log' to generate
an email body). But it shows that 'log -d' by itself can be useful.